AI model management holds the key to the stage of full-scale implementation
AI is getting closer and closer to the stage of full-scale implementation.
In terms of knowledge, not only does it continue to grow, but it also degenerates depending on environmental changes in the same way as a human brain.
A major challenge for the future is to establish an approach to AI model management that maintains optimal conditions for each area of use by making continuous improvements through repeated learning or by replacing the AI model itself.
Executive Officer, Principal
AI Sector Leader
P&T Digital Business Unit
Because it is a cutting-edge field, we need to confirm the maturity status of AI technology
These days, not a day goes by without the media reporting about artificial intelligence. In the business field, for example, there are reports about the launch of AI prototype verifications as well as implementation in various fields, such as development, production, marketing, and personnel.
However, these reports also include those in which it is difficult to determine whether a firm is using real AI or simply deploying existing technologies and referring to them as AI.
We have also heard about cases in which the use of AI has been abandoned because things haven’t gone well as intended, particularly following a rushed initial proof of concept (PoC), in order not to fall behind in the application of AI. In such cases, the cause of the failure may be attributed to beginning PoC without a specific idea of how to incorporate AI into the company’s business.
Because AI is an evolving field that has not yet matured, it is necessary to monitor the technical status of AI, following the most advanced research findings from international societies and scholars. On that basis, I believe that the starting point is to thoroughly ascertain in which areas of the company’s business and which business processes AI can be used, as well as values that can be created from that use.
A time will come when companies will use hundreds and thousands of new AI models
AI is formed and trained by letting algorithms learn from vast amount of data.
Therefore, AI grows in completely different ways, depending on the algorithms used and the data that is learned. Conversely, the optimal algorithms and learning data depends on the business processes to which the AI will be applied.
What we really need to keep in mind is the fact that every day, AI researchers around the world continue to research and publish unique ensembles of learning for new models that combine several new algorithms and models. Most of these models are publicized as theses and, therefore, are usable as open source materials.
Thus, from among those, companies should select the most suitable algorithm for the purpose of use, and train original AI models through data-based learning.
For example, a consumer goods manufacturer may want to use AI to make demand estimations for its products. Products can be divided into at least four categories, including new and existing products, each of which are divided into regular products and one-off products. Different variables affect the sales in each category, so it is possible to produce four AI models by selecting data sets and optimal algorithms for each of these.
Some companies may show their concerns saying “We don’t have resources to make and distribute multiple original AI models”.
Just one idea to resolve these concerns is a method of coordinating with AI platforms constructed in the cloud via API (Application Programming Interface).
This is the idea of using an API connection for AI that is beneficial because it is always up-to-date and has already been confirmed as an effective approach in the applicable area of business processes. In fact, we have already prepared such AI platforms on ABeam Cloud.
Today, it is possible to perform high-speed processing using vast amount of data; so it is not impossible to make more detailed classifications of each product category for distribution in an AI model, and to make highly accurate predictions. In the future, companies will be able to arrange an AI model for each product line and for each season, so that hundreds or even thousands of AI models will be used in total.
AI model management is a major challenge for companies
In the future, as companies become capable of generating and establishing various AI models, one major challenge will be the management of so many models. At ABeam Consulting, we call this “AI model management”.
However, no matter how highly precise the AI models formed, they will not be able to perform perpetually. As the environment in which they are placed changes, AI models will become unable to demonstrate their capabilities, and they will ultimately degenerate. In other words, environmental changes influence AI in the same way they influence humans.
For example, consider a case in which AI is used to predict the breakdown of manufacturing machinery. Now consider that the production line on which the machinery is placed changes. The same machinery, when placed in the factory in a different area, will change according to the temperature, humidity, and other aspects of the usage environment. The machinery operation status will change as well. That will cause a decline in the accuracy of breakdown prediction.
In other words, no matter how highly precise the AI models become, they will not be able to perform perpetually.
Therefore, it is necessary to monitor the performance of each AI model carefully, and if there is a decline in accuracy, the AI model must be improved through re-learning, and if it still does not achieve growth, subsequent measures must be taken to replace the AI model.
AI model management refers to the continuous repetition of the cycle of model construction, usage, evaluation, and re-learning (or replacement).
From an organizational perspective, another challenge is who should manage the AI model cycle. While there are various arguments in this regard, I personally believe that IT departments should manage systems related to business processes. Further, I believe that data science departments should manage AI related to business processes. Companies that have no data scientists need to either newly employ or train such people.
In the future, AI usage may expand to all areas of company business and business process fields. PoC should be carried out to discover the areas in which AI is to be implemented; and its projected effectiveness and return on investment should be determined, which will lead to the reform of business processes and system development. C-suite management executives, at the very least, must take responsibility for designing the road map to achieving that goal.
If AI usage is considered to be an important element in digital transformation, my theory at present is that the CDO (Chief Digital Officer) should take on that responsibility.